# Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

Source: [arXiv](https://arxiv.org/abs/2609.20822v1)  
Feed7 permalink: https://feed7.dev/p/2609-20822v1-1hfyx5i  
Published: 2026-09-17T17:59:58.000Z  
Trust: Needs Review (needs_review)

## Why Included

A safety prompt did not stop robot coding agents from colliding with obstacles. SafeHarness instead makes route verification, replanning, and contact constraints explicit in the execution loop.

## Source Summary

The tested coding agent recognized the obstacle and was explicitly told not to touch it, yet collided in most cases because safety never became a planning priority. SafeHarness reached **71.9% task success** and **87.5% collision avoidance**.

## Practical Implication

For safety-critical agents, encode constraints in the harness rather than relying on prompt compliance. Here that means grounding objects as boxes, generating and verifying waypoint routes, replanning blocked paths, and selecting a safe contact position.

## Agent-Ready Context

The tested coding agent recognized the obstacle and was explicitly told not to touch it, yet collided in most cases because safety never became a planning priority. SafeHarness reached **71.9% task success** and **87.5% collision avoidance**.

For safety-critical agents, encode constraints in the harness rather than relying on prompt compliance. Here that means grounding objects as boxes, generating and verifying waypoint routes, replanning blocked paths, and selecting a safe contact position.

The reported gains over the previous state of the art are **6.5%** for task success and **27.0%** for collision avoidance. The abstract does not establish whether the method transfers beyond its evaluated robot-manipulation tasks and obstacle setup.

## Connected Context

Feed7 judgment across 807 accumulated Signals:

This provides task-level evidence that explicit safety instructions can fail even when the agent recognizes the hazard. It strengthens the case for moving constraints from prompts into executable planning and verification, and makes that principle concrete through grounded geometry and checked routes. Transfer beyond the tested manipulation setup remains unresolved.

- [In Code They Act, In Proof We Trust — Erik Meijer, Leibniz Labs](https://feed7.dev/p/in-code-they-act-in-proof-we-trust-erik-meijer-leibniz-labs-1lyno2y) — SafeHarness operationalizes the proposed proof-before-side-effects pattern in robotics by verifying waypoint routes before physical execution.
- [Twin: Playing an Unknown Game with a Test-Time Digital Twin](https://feed7.dev/p/2608-14490v1-0d3xjvt) — Both gate real actions with an executable domain model; Twin validates simulated transitions, whereas SafeHarness verifies geometric routes and replans around obstacles.
- [The Design-Code Roundtrip That Isn't — Jonathan Gordon, ReWeaver AI](https://feed7.dev/p/the-design-code-roundtrip-that-isn-t-jonathan-gordon-reweaver-ai-0fkffut) — Both show deterministic checks catching risks that iterative generation can introduce or ignore, though this Signal moves the gate from artifact comparison to physical collision avoidance.

## Context Map

- Layer: agent
- Domains: coding
- Topics: harness-engineering, agent-reliability

## Uncertainty

- The reported gains over the previous state of the art are **6.5%** for task success and **27.0%** for collision avoidance. The abstract does not establish whether the method transfers beyond its evaluated robot-manipulation tasks and obstacle setup.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.